A large structure dynamic response prediction method based on a hybrid deep learning model

Abstract

The present application relates to a kind of large structure dynamic response prediction method based on hybrid deep learning model, comprising: the finite element model of large structure is established;Modal parameters of large structure are measured using exciter, and the finite element model of large structure is corrected according to modal parameters of large structure;The dynamic response of large structure under dynamic load is simulated using the finite element model after correction;While using sensor actual collection dynamic response of large structure under dynamic load;The dynamic response of large structure under dynamic load is simulated and actually collected by simulation and construction training set;According to the training set constructed, hybrid deep learning model is trained;Dynamic load acting on large structure is input into trained deep learning model to predict the dynamic response of large structure.The present application improves the prediction efficiency and accuracy, enhances the generalization ability of model, so that it can adapt to various different environments and conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of structural dynamic response prediction, and particularly relates to a large-scale structure dynamic response prediction method based on a hybrid deep learning model. BACKGROUND

[0002] With the rapid development of modern industry and cities, large structures such as skyscrapers, bridges, power transmission towers, dams, and industrial facilities have become key pillars of modern social infrastructure and technological progress. These structures not only play a crucial role in economic development, supporting the efficient operation of multiple fields such as transportation, energy, and communication, but also play an indispensable role in ensuring public safety, providing employment opportunities, and promoting technological innovation. The stability and durability of large structures are directly related to human well-being and environmental sustainability, therefore, in-depth study of the dynamic response of large structures under dynamic loads is crucial to ensure their safe operation.

[0003] In the field of dynamic response prediction of large structures, traditional prediction methods face many challenges. On the one hand, relying solely on finite element models for dynamic response prediction, due to the complexity of actual structures and the approximation process in modeling, the model often fails to accurately reflect the behavior of large structures under real dynamic loads. Factors such as material properties, boundary conditions, and other uncertainties of actual structures lead to deviations between finite element model prediction results and actual conditions.

[0004] On the other hand, relying solely on sensor-collected data for prediction, while it can reflect the actual response of the structure, lacks a deep understanding of the internal physical mechanisms of the structure, and the data is often affected by factors such as noise and sensor accuracy.

[0005] Existing deep learning models also have limitations when dealing with such problems. A single deep learning model, such as a convolutional neural network (CNN), may not be able to fully capture feature information at different scales when dealing with large-scale structure dynamic response data with complex spatiotemporal characteristics; long short-term memory networks (LSTM) have good processing capabilities for time series data, but lack richness in feature extraction. Therefore, a method that can consider the advantages of both structural physical models and deep learning is needed to improve the accuracy and reliability of large-scale structure dynamic response prediction. SUMMARY

[0006] To solve the problems in the background art, the present application provides a large-scale structure dynamic response prediction method based on a hybrid deep learning model, comprising:

[0007] S1: Establish a finite element model of a large structure;

[0008] S2: measuring modal parameters of the large structure by using the exciter, and correcting the finite element model of the large structure according to the modal parameters of the large structure;

[0009] S3: simulating the dynamic response of the large structure under dynamic load by using the corrected finite element model; and collecting the dynamic response of the large structure under dynamic load by using the sensor;

[0010] S4: constructing a training set by simulating and collecting the dynamic response of the large structure under dynamic load;

[0011] S5: training a hybrid deep learning model according to the constructed training set, wherein the hybrid deep learning model comprises a multi-scale convolutional neural network MS-CNN, a multi-head attention mechanism Multi-Head Attention and a long short-term memory network LSTM; the multi-scale convolutional neural network MS-CNN is used for multi-scale feature extraction of the input training sample, the multi-head attention mechanism Multi-Head Attention is used for attention enhancement processing of the extracted multi-scale features, and the long short-term memory network LSTM is used for predicting the dynamic response of the large structure according to the attention-enhanced features;

[0012] S6: inputting the dynamic load acting on the large structure into the trained deep learning model to predict the dynamic response of the large structure.

[0013] The present application has at least the following beneficial effects

[0014] The present application measures the modal parameters of a large structure by using a vibration exciter, and corrects the finite element model accordingly, so that the finite element model can more accurately reflect the actual characteristics of the large structure, providing a more reliable basis for subsequent simulation, effectively reducing the prediction deviation caused by model error. The dynamic response of the large structure under dynamic load is simulated using the corrected finite element model, and the response data is actually collected using sensors; and a training set is constructed based on the two kinds of data. This combination of simulation and measured data makes full use of the advantages of data from different sources, makes up for the shortcomings of a single data source, enriches the information of the training data, and helps to improve the generalization ability and prediction accuracy of the model. The present application adopts a hybrid deep learning model containing a multi-scale convolutional neural network (MS-CNN), a multi-head attention mechanism (Multi-Head Attention) and a long short-term memory network (LSTM). The MS-CNN can extract multi-scale features from the input training samples, capture the structural dynamic response features at different scales, and enrich the feature information. The multi-head attention mechanism enhances the attention of the extracted multi-scale features, so that the model can pay more attention to key features and improve the expression ability of the features. The LSTM predicts the dynamic response of the large structure based on the attention-enhanced features, fully utilizes its processing capability for time series data, effectively fuses different types of feature information, and improves the accuracy and reliability of the prediction of the dynamic response of the large structure. It provides a convenient and efficient prediction method for engineering practical applications, which helps engineers to understand the response of the structure under different load conditions in a timely manner, and provides strong support for structural design, maintenance and safety evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The method flowchart of the present application is shown in the figure;

[0016] Figure 2 The dynamic model of the power transmission tower in the embodiment of the present application is shown in the figure;

[0017] Figure 3 The wind speed time history graph and the wind power spectrum comparison graph in the embodiment of the present application are shown in the figure;

[0018] Figure 4 The network structure diagram of the deep learning hybrid model in the embodiment of the present application is shown in the figure;

[0019] Figure 5 The prediction graph of the dynamic response (displacement response) of the power transmission tower in the embodiment of the present application is shown in the figure;

[0020] Figure 6 The prediction graph of the dynamic response (stress response) of the power transmission tower in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0021] Following, the embodiments of the present application will be described in detail by specific examples, and other advantages and effects of the present application can be easily understood by those skilled in the art from the disclosure of this specification. The present application can also be implemented or applied by other different embodiments, and various modifications or changes can be made to the details in this specification based on different views and applications without departing from the spirit of the present application. It should be noted that the drawings provided in the following examples only illustrate the basic concept of the present application in a schematic manner, and the following examples and features in the examples can be combined with each other without conflict.

[0022] Referring to Figure 1 The present application provides a large structure dynamic response prediction method based on a hybrid deep learning model, comprising:

[0023] S1: establishing a finite element model of the large structure;

[0024] Preferably, the step of establishing the finite element model of the large structure comprises: constructing a three-dimensional model of each component of the large structure according to the design drawings of the large structure in SolidWorks, importing the three-dimensional model of the large structure into Abaqus after completing the SolidWorks modeling, and establishing the finite element model of the large structure by dividing the grid, defining the material properties, and setting the contact and constraint conditions.

[0025] Referring to Figure 2 Taking a power transmission tower as an example, in this embodiment, a scaled physical model of the power transmission tower is established according to the detailed design drawings of the power transmission tower, then the numerical model of each component (tower body, tower arm, connecting piece, etc.) of the power transmission tower is created one by one according to the design drawings by using Solid Works software. After completing the modeling of all components, they are imported into Abaqus, a professional finite element analysis software, and the finite element model of the structure is constructed by dividing the grid, defining the material properties, setting the contact and constraint conditions, etc.

[0026] S2: measuring the modal parameters of the large structure by using an exciter, and correcting the finite element model of the large structure according to the modal parameters of the large structure;

[0027] Preferably, the step S2 comprises:

[0028] S21: applying a dynamic load to the large structure by using an exciter; the dynamic load comprises an excitation force with a known frequency and amplitude;

[0029] In the field of the power transmission tower, an exciter is installed. The exciter is a device capable of generating an excitation force with a specific frequency and amplitude. According to the pre-set experimental scheme, a suitable exciter is selected and fixed on a specific position of the power transmission tower, such as a certain standard section of the tower body.

[0030] In order to comprehensively obtain the modal parameters of the power transmission tower, the exciter is set to generate dynamic loads of different frequencies and amplitudes. For example, first set the exciter to generate an excitation force with a frequency of 5 Hz and an amplitude of 100 N to apply a dynamic load to the power transmission tower. This excitation force will cause the power transmission tower to vibrate.

[0031] S22: measuring the dynamic response of the large structure using sensors, measuring the modal parameters of the large structure through the sensors, the modal parameters of the large structure including: the vibration mode of the large structure under different loads, the natural frequency of vibration and the decay rate of vibration;

[0032] A plurality of sensors, such as acceleration sensors and displacement sensors, are installed at key positions of the power transmission tower.

[0033] After the exciter applies the dynamic load, the power transmission tower begins to vibrate, and the sensors measure the dynamic response of the tower in real time. For example, the acceleration sensor measures the acceleration change of the tower during vibration, and the displacement sensor measures the displacement change of each part of the tower.

[0034] By analyzing and processing the data collected by the sensors, the modal parameters of the power transmission tower can be obtained. Specifically, by analyzing the characteristics of the vibration signal, the vibration mode of the tower under different loads is identified, such as whether it is overall vibration or local vibration, the direction and amplitude of vibration, etc.; the natural frequency of vibration is determined through frequency spectrum analysis, i.e. the vibration frequency of the tower itself when it is free to vibrate; the decay rate of vibration is calculated by observing the decay of the vibration signal over time.

[0035] S23: Adjust the grid division, define material properties, set contact and constraint condition parameters of the finite element model, so that the modal parameters of the large structure measured and identified by the sensors are consistent with the theoretical modal parameters calculated by the finite element model within the preset error range, and the finite element model is modified.

[0036] In this embodiment, the theoretical modal parameters calculated by the previously established finite element model of the transmission tower are viewed in the finite element model analysis software Abaqus. The modal parameters identified by the sensor measurement are compared with the theoretical modal parameters calculated by the finite element model. If a large difference is found between the two, the number of grids is appropriately increased or decreased, and the shape and size distribution of the grids are changed. It is checked whether the previously input material parameters are accurate. If there is a certain deviation in the elastic modulus of the steel actually used, the elastic modulus and other material property parameters are adjusted again, and the modal parameters of the finite element model are calculated again to observe the closeness to the actual measured values. It is checked whether the contact settings between the tower sections are reasonable and whether the contact stiffness of the bolt connection conforms to the actual situation. For example, the contact stiffness between the bolt and the connecting plate is appropriately adjusted, or the constraint conditions of the tower bottom are modified to make them more consistent with the actual installation and stress state, and then the modal parameters are recalculated. By continuously adjusting these parameters, the theoretical modal parameters calculated by the finite element model are consistent with the actual modal parameters identified by the sensor measurement within a preset error range.

[0037] S3: simulating and simulating the dynamic response of the large structure under the dynamic load by using the corrected finite element model; and simultaneously collecting the dynamic response of the large structure under the dynamic load by using the sensor;

[0038] Preferably, the dynamic response data includes one or more of the displacement, stress and strain of the measurement point on the large structure.

[0039] S4: constructing a training set by simulating and collecting the dynamic response of the large structure under the dynamic load;

[0040] Preferably, the training samples in the training set include the dynamic load V acting on the large structure and the corresponding dynamic response U.

[0041]

[0042] wherein v ti represents the dynamic load of the large structure at t i ; u ti represents the dynamic response of the dangerous area of the large structure at t i .

[0043] Specifically, in this embodiment, the wind-induced dynamic response data of the transmission tower obtained by simulation analysis and experiment are preprocessed to construct a data set required by a deep learning hybrid model. The data set is divided into a training set, a validation set and a test set according to a ratio of 6:3:3, wherein the ratio of real data to simulation data in the training set is 4:2, and the ratio of real data to simulation data in the validation set and the test set is 2:1.

[0044] Please refer to Figure 4S5: training the hybrid deep learning model according to the constructed training set, wherein the hybrid deep learning model comprises a multi-scale convolutional neural network MS-CNN, a multi-head attention mechanism Multi-Head Attention, and a long short-term memory network LSTM; the multi-scale convolutional neural network MS-CNN is configured to perform multi-scale feature extraction on the input training sample, the multi-head attention mechanism Multi-Head Attention is configured to perform attention enhancement processing on the extracted multi-scale features, and the long short-term memory network LSTM is configured to predict the dynamic response of the large structure according to the attention-enhanced features;

[0045] Preferably, the multi-scale convolutional neural network MS-CNN comprises four parallel convolutional paths, each of which uses a different size of convolution kernel and a different expansion coefficient to capture multi-scale information of the input features; the four parallel convolutional paths are configured to perform four-scale feature extraction on the dynamic load V acting on the large structure to obtain four-scale features T×P×D , T is a time dimension, P is a position dimension, and D is a feature dimension. and The output features of each convolutional path are spliced in the channel dimension to obtain multi-scale features F MS ∈R C×H×W , wherein C = C1 + C2 + C3 + C4, C represents the channel dimension of the multi-scale features F MS , C j represents the channel dimension of the features F j , and j ∈ {1, 2, 3, 4}.

[0046] In the embodiment, the multi-scale feature extraction by the four parallel convolutional paths of the multi-scale convolutional neural network MS-CNN comprises:

[0047] The first convolutional path: using a 1×1 convolution kernel to extract features from the input, obtaining features F1 with a dimension of [C1, H, W], the 1×1 convolution is used for dimension reduction and feature channel compression, while retaining the local feature details.

[0048] wherein C1 is the number of channels;

[0049]

[0050] K is the size of the convolution kernel, d is the expansion coefficient, P is the size of the edge zero padding, and S is the step size.

[0051] Second convolution path: 3x3 and 1x1 convolution kernels are used in turn to extract features from the input. The expansion coefficient of the first convolution layer is 1, which is used to capture medium-scale local features; the second 1x1 convolution is used to further adjust the feature channel number. Each convolution layer adopts the edge zero padding method to ensure that the spatial size of the feature map is the same as F1, and finally the feature F2 with a dimension of [C2, H, W] is obtained.

[0052] Third convolution path: 5x5 and 1x1 convolution kernels are used in turn to extract features from the input. The expansion coefficient of the first convolution layer is 2, which is used to expand the receptive field and capture larger-scale feature information; the second 1x1 convolution is used to adjust the feature channel. Similarly, the edge zero padding method is adopted to obtain the feature F3 with a dimension of [C3, H, W].

[0053] Fourth convolution path: 1x1 and 3x3 convolution kernels are used in turn to extract features from the input. The first 1x1 convolution is used for dimension reduction, and the second 3x3 convolution layer has an expansion coefficient of 1, which is used to extract local features. Through edge zero padding, the final feature F4 with a dimension of [C4, H, W] is obtained. The data input adopts the method of multi-scale convolution, that is, by using different sizes of convolution kernels and the method of empty convolution, the result matrix processed by different convolution layers in the depth dimension is spliced to form a deeper matrix. Smaller convolution kernels focus on the local details of the input data and extract fine feature information; while larger convolution kernels use the method of empty convolution to capture broader regional features. The information aggregation on different sizes is realized, and the features are extracted from different scales, making the model more flexible and robust when dealing with complex scenes.

[0054] In order to further optimize the calculation efficiency and improve the performance of the model, the multi-head attention mechanism is used to accurately screen out the key features closely related to the output, which effectively reduces the data dimension and reduces the redundant information. Subsequently, the generated multi-dimensional tensor is flattened into a one-dimensional tensor, which is input into the long short-term memory network (LSTM). The cyclic structure of LSTM can effectively capture the time-dependent relationship in the data, thereby effectively integrating the input features and predicting or outputting the corresponding response, and finally outputting the dynamic response result of the power tower.

[0055] Preferably, the multi-head attention mechanism Multi-Head Attention includes:

[0056] Feature mapping: the multi-scale features F MS ∈R C×H×W Apply three parallel 1x1 convolution operations to generate features F5∈R C×H×W , F6∈R C×H×W and F7∈R C×H×Wcharacteristics F5, F6 and F7 are respectively taken as a query vector Q, a key vector K and a value vector V in a multi-head self-attention mechanism;

[0057] Multi-head self-attention calculation: the query vector Q, the key vector K and the value vector V are evenly divided into M groups in the channel dimension, and the number of channels in each group is C / M, to obtain the grouped query vector Q m , the key vector K m and the value vector V m , wherein m={1, 2, …, m, …, M};

[0058] For each group of query vectors Q m , key vectors K m and value vectors V m , the attention enhanced feature Z m is calculated through the self-attention mechanism:

[0059]

[0060] wherein d k is a temperature coefficient, and softmax represents an activation function;

[0061] Multi-head feature splicing: the attention enhanced features output by the M groups of self-attention are spliced in the channel dimension to obtain the multi-head attention mechanism Multi-Head Attention attention enhanced feature F att ∈R C×H×W =concat(Z1,Z2,...,Z m ,...,Z M ), and concat represents a splicing function.

[0062] Preferably, the long short-term memory network LSTM comprises:

[0063] Input feature processing: the input feature F att is flattened from the shape [C, H, W] to [C, HxW], and the spatial dimension is converted into sequence data through the flattening operation, forming a feature sequence with a time step of HxW and a feature dimension of C;

[0064] Time series modeling and feature integration: a two-layer LSTM unit is used to model the time series and integrate the features of the flattened feature sequence, the input dimension of the LSTM unit is C, and the hidden layer size is hidden-size, the long-term dependency relationship between the features is captured through the LSTM unit, and the output shape of the time series feature is [hidden-size, HxW];

[0065] Feature mapping and prediction: the output of the last time step of the LSTM is mapped to the target dimension [S] through a fully connected layer, where S is the dimension of the dynamic response data required for prediction, to realize the dynamic response prediction of large structures.

[0066] S6: input the dynamic load acting on the large structure into the trained deep learning model to predict the dynamic response of the large structure.

[0067] Preferably, the loss function used when training the hybrid deep learning model is:

[0068]

[0069] where L MSE represents the loss function, m is the number of samples, u predict is the predicted data, u target is the true data.

[0070] Figure 2 The elevation view of the wine glass type transmission tower is given, and the total height of the tower is 42.8m. Taking this tower as an example for analysis. Based on the davenport wind speed power spectrum, the harmonic superposition method is used to simulate the wind speed. For a scaled model of a transmission tower, wind tunnel tests are used to load the wind speed, and dynamic response sensors are installed on the transmission tower in proportion. The sampling frequency is set to 1Hz to collect displacement and stress data of the transmission tower under wind load; for the finite element model, the wind speed is converted to wind pressure according to the transmission tower design manual and input into the finite element model for dynamic response analysis. Figure 3 The wind speed time history curve is given, and compared with the davenport fluctuating wind speed power spectrum, it is verified that the simulated wind speed time history curve meets the requirements. The trained weights are loaded into the deep learning hybrid model, and the dynamic load data measured by the sensor is preprocessed and put into the trained deep learning model to predict the dynamic response of the transmission tower. The comparison chart of the time history curve of the maximum displacement and the maximum stress position of the transmission tower and the predicted value is given, as shown in Figure 5 and Figure 6 The accuracy of the calculated displacement and stress response is 98.87% and 97.65% respectively. At the same time, the performance of the model in terms of prediction accuracy, noise robustness and calculation efficiency is evaluated.

[0071] In summary, the present application measures the modal parameters of the large structure by using the exciter, and corrects the finite element model accordingly, so that the finite element model can more accurately reflect the actual characteristics of the large structure, and provide a more reliable basis for subsequent simulation, effectively reducing the prediction deviation caused by model error. The dynamic response of the large structure under dynamic load is simulated using the corrected finite element model, and the response data is actually collected using sensors; and a training set is constructed based on the two kinds of data. This combination of simulation and measured data makes full use of the advantages of data from different sources, makes up for the shortcomings of a single source of data, enriches the information of the training data, and helps to improve the generalization ability and prediction accuracy of the model. The present application adopts a hybrid deep learning model containing a multi-scale convolutional neural network (MS-CNN), a multi-head attention mechanism (Multi-Head Attention) and a long short-term memory network (LSTM). The MS-CNN can extract multi-scale features from the input training samples, capture the structural dynamic response features at different scales, and enrich the feature information. The multi-head attention mechanism performs attention enhancement processing on the extracted multi-scale features, so that the model can pay more attention to key features and improve the expression ability of the features. The LSTM predicts the dynamic response of the large structure based on the attention-enhanced features, fully utilizes its processing capability for time series data, effectively fuses different types of feature information, and improves the accuracy and reliability of the prediction of the dynamic response of the large structure. It provides a convenient and efficient prediction method for engineering practical applications, and helps engineers to understand the response of the structure under different load conditions in a timely manner, and provides strong support for structural design, maintenance and safety evaluation.

[0072] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting dynamic response of large-scale structures based on a hybrid deep learning model, characterized in that, The method comprises the following steps: S1: establishing a finite element model of the large structure; S2: measuring the modal parameters of the large structure by using an exciter, and correcting the finite element model of the large structure according to the modal parameters of the large structure; The step S2 comprises: S21: applying a dynamic load to the large structure by using an exciter; the dynamic load comprises an excitation force with a known frequency and amplitude; S22: measuring the dynamic response of the large structure by using a sensor, and measuring the modal parameters of the large structure by using the sensor; the modal parameters of the large structure comprise the vibration mode, the natural frequency and the decay speed of the vibration of the large structure under different loads; S23: correcting the finite element model by adjusting the mesh division, defining the material properties, setting the contact and constraint condition parameters, so that the modal parameters of the large structure measured by the sensor are consistent with the theoretical modal parameters calculated by the finite element model within a preset error range; The dynamic response data comprises one or more of the displacement, stress and strain of the measuring point on the large structure; S3: simulating the dynamic response of the large structure under the dynamic load by using the corrected finite element model; and simultaneously collecting the dynamic response of the large structure under the dynamic load by using the sensor; S4: constructing a training set by simulating and collecting the dynamic response of the large structure under the dynamic load; S5: training a hybrid deep learning model according to the constructed training set; wherein the hybrid deep learning model comprises a multi-scale convolutional neural network MS-CNN, a multi-head attention mechanism Multi-Head Attention and a long short-term memory network LSTM; the multi-scale convolutional neural network MS-CNN is used for multi-scale feature extraction of the input training sample, the multi-head attention mechanism Multi-Head Attention is used for attention enhancement processing of the extracted multi-scale features, and the long short-term memory network LSTM is used for predicting the dynamic response of the large structure according to the attention-enhanced features; The loss function used for training the hybrid deep learning model is: where L MSE denotes the loss function, m is the number of samples, u predict is the predicted data, u target is the true data; S6: inputting the dynamic load acting on the large structure into the trained deep learning model to predict the dynamic response of the large structure.

2. The method of claim 1, wherein the method is based on a hybrid deep learning model. The step of establishing the finite element model of the large structure comprises: constructing a three-dimensional model of each part of the large structure in SolidWorks according to the design drawings of the large structure; after completing the modeling in SolidWorks, importing the three-dimensional model of the large structure into Abaqus, and establishing the finite element model of the large structure by dividing the mesh, defining the material properties, and setting the contact and constraint conditions.

3. The method of claim 1, wherein the method is characterized by, The training samples in the training set comprise a dynamic load V acting on the large structure and a corresponding dynamic response U; where v ti represents the dynamic load of the large-scale structure at t i ; u ti represents the dynamic response of the dangerous area of the large-scale structure at t i .

4. The method of claim 1, wherein the method is based on a hybrid deep learning model. The multi-scale convolutional neural network MS-CNN comprises four parallel convolutional channels, each of which adopts a different size of convolution kernel and a different expansion coefficient to capture multi-scale information of input features; the four parallel convolutional channels are used to perform four-scale feature extraction on dynamic load V acting on a large structure, V∈R T×P×D , T is a time dimension, P is a position dimension, and D is a feature dimension and The output features of each convolutional channel are spliced in the channel dimension to obtain multi-scale features F MS ∈R C×H×W , wherein C=C1+C2+C3+C4, C represents a channel dimension of the multi-scale features F MS , C j represents a channel dimension of the features F j , and j∈{1, 2, 3, 4}.

5. The method of claim 4, wherein the method is characterized by, The multi-head attention mechanism Multi-Head Attention comprises: Feature mapping: multi-scale features F MS ∈R C×H×W Apply three parallel 1x1 convolution operations to generate features F5∈R C×H×W , F6∈R C×H×W and F7∈R C×H×W , respectively, as the query vector Q, the key vector K and the value vector V in the multi-head self-attention mechanism; Multi-head self-attention calculation: the query vector Q, the key vector K and the value vector V are uniformly divided into M groups in the channel dimension, and the number of channels in each group is C / M, to obtain the grouped query vector Q m , the grouped key vector K m and the grouped value vector V m , wherein m={1, 2, …, m, …, M}. For each group of query vectors Q m , key vectors K m and value vectors V m , attention enhanced features Z m are computed by a self-attention mechanism. where d k is a temperature coefficient, and softmax denotes an activation function. Multi-head feature concatenation: the attention enhanced features output by the M groups of self-attention are concatenated in the channel dimension to obtain the features F after attention enhancement of the multi-head attention mechanism Multi-Head Attention att ∈R C×H×W = concat(Z1, Z2,..., Z m ,..., Z M ), concat denotes a concatenation function.

6. The method of claim 5, wherein the method is characterized by, The long short-term memory network LSTM The method comprises the following steps: Input feature processing: the input feature F att Flatten from shape [C, H, W] to [C, HxW], convert the spatial dimension to sequence data by the flatten operation, form a feature sequence with time step HxW and feature dimension C; Temporal modeling and feature integration: use two layers of LSTM units to model the temporal sequence and integrate the features of the flattened feature sequence, the input dimension of the LSTM unit is C, the hidden layer size is hidden-size, and the long-term dependence between features is captured through the LSTM unit, and the output shape is [hidden-size, HxW] of the temporal feature; Feature mapping and prediction: the output of the last time step of the LSTM is mapped to the target dimension [S] through a fully connected layer, where S is the dimension of the required predicted dynamic response data, and the dynamic response prediction of large structures is realized.

Images